# CVE-2026-111392: SGLang cache poisoning via untrusted mm_hashes in multimodal prefix cache

Source: https://junglewise.ai/threats/cve-2026-111392-sglang-through-0-5-21-trusts-caller-supplied-mm-hashes-values-as  
Publisher: Junglewise Threat Intelligence, https://junglewise.ai/threats  
Published: 2026-10-16. Updated: 2026-10-16  
License: CC BY 4.0 with a link back

CVE-2026-111392 · Severity: medium · CVSS 4.8 · Published 2026-10-16

SGLang is a language model serving runtime that caches preprocessed multimodal data (images) to improve request performance. The vulnerability allows remote attackers to supply fake hash values for cached image data, causing the server to return a different image than originally cached. An attacker can poison the cache to disclose images from other users' requests or corrupt responses.

Technologies: [SGLang Project SGLang](https://junglewise.ai/threats/technologies/sglang). Vendors: [SGLang Project](https://junglewise.ai/threats/vendors/sglang-project).

## Quick answers

**Is CVE-2026-111392 exploited in the wild?** No. Junglewise Threat Intelligence has no report of CVE-2026-111392 being exploited in the wild.

**How severe is CVE-2026-111392?** Medium severity, with a CVSS score of 4.8, published 16 October 2026.

**Which products and versions are affected?** SGLang Project SGLang through 0.5.21.

**Is there a fix for CVE-2026-111392?** No fixed version is recorded yet. Check the vendor's advisory in the references below. If you run it, Junglewise drafts a mitigation that fits your setup and applies it once you approve, with senior security engineers behind it.

**How is CVE-2026-111392 attacked?** It is attacked over the network, in a software library.

**Am I affected by CVE-2026-111392?** Check whether you run SGLang Project SGLang at an affected version. Junglewise does that check for you around the clock: connect your code, cloud and domains, and it matches new vulnerabilities like this one against what you actually run and tells you when one reaches you.

## Technical details

SGLang through version 0.5.21 uses caller-supplied mm_hashes values as the key for multimodal prefix-cache lookups without validation. An attacker can send a request via the /generate endpoint with a forged mm_hash matching a known or predicted hash from another request, paired with a different image, to corrupt cached KV entries. This is a cache poisoning vulnerability requiring network access but no authentication.

## Affected products

- SGLang Project SGLang through 0.5.21

## Timeline

- 2026-10-16: disclosed

## References

- [https://github.com/sgl-project/sglang/blob/v0.5.21/python/sglang/srt/managers/io_struct.py](https://github.com/sgl-project/sglang/blob/v0.5.21/python/sglang/srt/managers/io_struct.py)
- [https://github.com/sgl-project/sglang/blob/v0.5.21/python/sglang/srt/managers/tokenizer_manager.py](https://github.com/sgl-project/sglang/blob/v0.5.21/python/sglang/srt/managers/tokenizer_manager.py)

## Related threats

- [SGLang uncontrolled memory allocation in Mooncake disaggregation backend](https://junglewise.ai/threats/cve-2026-112270-sglang-through-0-5-22-with-the-mooncake-disaggregation-backend)
- [SGLang Mooncake disaggregation backend unauthenticated denial of service](https://junglewise.ai/threats/cve-2026-112269-sglang-through-0-5-22-with-the-mooncake-disaggregation-backend)
- [SGLang server-side request forgery in prefill/decode disaggregation](https://junglewise.ai/threats/cve-2026-112267-sglang-through-0-5-22-in-prefill-decode-disaggregation-mode)
- [SGLang code injection in _evaluate_filter()](https://junglewise.ai/threats/cve-2026-111391-sglang-0-5-10-through-0-5-22-contains-a-code-injection)
- [SGLang denial of service via disabled image decompression protection](https://junglewise.ai/threats/cve-2026-109229-sglang-through-0-5-21-contains-a-denial-of-service-vulnerability)